Authentication method and apparatus, and method and apparatus for training a recognizer

• US 10,275,684 B2
• Filed: 10/18/2016
• Issued: 04/30/2019
• Est. Priority Date: 11/04/2015
• Status: Active Grant

• Current Assignee: Samsung Electronics Co. Ltd.
• Original Assignee: Samsung Electronics Co. Ltd.
• Inventors: Han, Seungju, Yoo, Byungin, Han, Jaejoon
• Primary Examiner(s): Chen, Xuemei G
• Application Number:
  US15/296,503
• Publication Number:
  US 20170124428A1
• Time in Patent Office: 924 Days
• Field of Search: 348 E7081, 348142, 348169, 382103, 382117, 382157, 382224, 600408, 706 20
• CPC Class Codes: G06F 18/22 Matching criteria, e.g. proximity measures
• CPC Class Codes: G06F 18/285 Selection of pattern recognition techniques, e.g. of classifiers in a multi-classifier system
• CPC Class Codes: G06F 21/32 using biometric data, e.g. fingerprints, iris scans or voiceprints
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06V 10/454 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
• CPC Class Codes: G06V 10/761 Proximity, similarity or dissimilarity measures
• CPC Class Codes: G06V 10/82 using neural networks
• CPC Class Codes: G06V 40/172 Classification, e.g. identification

Abstract

An authentication method and apparatus, and a method and apparatus for training a recognizer are provided. The authentication method may include determining whether authentication of an input image succeeds by comparing a similarity associated with an intermediate layer to a threshold.

Claims

1. An authentication method comprising:

applying an input image to a recognizer, the recognizer including a plurality of layers of a deep learning (DL) network, the input image including a first object; and

determining whether the first object is the same as a second object in a registered image, the determining including,

determining a similarity or a distance between at least one feature value of the input image and at least one feature value of the registered image corresponding to a first layer of the layers in the recognizer, the at least one feature value of the input image output by the first layer of the layers in the recognizer,

determining whether the determined similarity or distance is within a threshold range corresponding to the first layer of the layers in the recognizer, the threshold range corresponding to the first layer of the layers in the recognizer being different than a threshold range corresponding to a second layer of the layers in the recognizer, and

authenticating the input image based on whether the similarity or the distance is within the threshold range corresponding to the first layer of the layers in the recognizer.

2. The authentication method of claim 1, wherein the authenticating comprises:

determining whether the input image is accepted or rejected.

3. The authentication method of claim 1, wherein the first layer belongs to a plurality of intermediate layers, and the determining whether the first object is the same as the second object in the registered image comprises:

determining whether the authenticating succeeds based on feature values of the input image and feature values of the registered image, the feature values of the input image output by the intermediate layers, respectively, and the feature values of the registered image corresponding to the intermediate layers, respectively.

4. The authentication method of claim 1, wherein the determining whether the first object is the same as the second object in the registered image comprises:

determining the similarity between the at least one feature value of the input image and the at least one feature value of the registered image, the authenticating being based on the determined similarity.

5. The authentication method of claim 4, wherein the authenticating comprises:

determining that the input image is accepted when the determined similarity exceeds a similarity corresponding to a false acceptance rate (FAR) of the first layer.

6. The authentication method of claim 5, wherein the authenticating comprises:

determining that the input image is rejected when the determined similarity is less than a similarity corresponding to a verification rate (VR) of the first layer, the similarity corresponding to the FAR and the similarity corresponding to the VR being different.

7. The authentication method of claim 6, wherein the determining whether the first object is the same as the second object in the registered image comprises:

proceeding to a next layer of the recognizer when the determined similarity exceeds the similarity corresponding to the VR and is less than the similarity corresponding to the FAR.

8. The authentication method of claim 4, wherein the threshold range comprises a first threshold part based on a false acceptance rate (FAR) and a second threshold part based on a verification rate (VR).

9. The authentication method of claim 8, wherein the first threshold part is greater than the second threshold part.

10. The authentication method of claim 1, wherein the determining whether the first object is the same as the second object in the registered image comprises:

determining the distance between the at least one feature value of the input image and the at least one feature value of the registered image, the authenticating being based on the determined distance.

11. The authentication method of claim 10, wherein the authenticating comprises:

determining that the input image is accepted when the determined distance is less than a distance corresponding to a false acceptance rate (FAR) of the first layer.

12. The authentication method of claim 11, wherein the authenticating comprises:

determining that the input image is rejected when the determined distance exceeds a distance corresponding to a verification rate (VR) of the first layer.

13. The authentication method of claim 12, wherein the determining whether the first object is the same as the second object in the registered image comprises:

proceeding to a next layer of the recognizer when the determined distance exceeds the distance corresponding to the FAR and is less than the distance corresponding to the VR.

14. The authentication method of claim 10, wherein the threshold range comprises a first threshold part based on a false acceptance rate (FAR) and a second threshold part based on a verification rate (VR).

15. The authentication method of claim 14, wherein the first threshold part is greater than the second threshold part.

16. The authentication method of claim 1, wherein the determining whether the first object is same as the second object in the registered image further includes:

determining whether the authenticating succeeds based on a final layer feature value of the input image and a final layer feature value of the registered image, the final layer feature value output by a final layer in the recognizer and the final layer feature value of the registered image corresponding to the final layer.

17. The authentication method of claim 1, further comprising:

applying the registered image to the recognizer; and

storing at least one feature value output by at least one intermediate layer and a final layer feature value output by a final layer of the plurality of layers as feature values of the registered image.

18. The authentication method of claim 17, further comprising:

receiving the registered image from a camera.

19. The authentication method of claim 1, wherein the first layer belongs to a plurality of intermediate layers among the plurality of layers.

20. The authentication method of claim 1, wherein the plurality of layers comprise at least one convolution layer and at least one pooling layer.

21. A training method comprising:

training a recognizer, the recognizer including a plurality of layers of a deep learning (DL) network;

determining a first threshold range for a first layer in the trained recognizer based on a desired performance index, the first threshold range being used for early authentication of an input image, the early authentication being a termination of an authentication method before reaching a final layer of the plurality of layers in the trained recognizer based on the first threshold range and authentication of the input image; and

determining a second threshold range for a second layer in the trained recognizer based on the desired performance index.

22. The training method of claim 21, further comprising:

applying validation data to the trained recognizer;

determining a first similarity between first feature values, the first feature values output by the first layer;

determining a second similarity between second feature values, the second feature values output by the second layer;

determining the first threshold range based on the first similarity; and

determining the second threshold range based on the second similarity.

23. The training method of claim 21, further comprising:

applying validation data to the trained recognizer;

determining a first distance between first feature values, the first feature values output by the first layer;

determining a second distance between second feature values, the second feature values output by the second layer;

determining the first threshold range based on the first distance; and

determining the second threshold range based on the second distance.

24. The training method of claim 21, wherein the first threshold range and the second threshold range are used to determine whether the input image is accepted or rejected.

25. The training method of claim 21, wherein the determining of the first threshold range comprises:

determining the first threshold range based on a similarity between a first feature value of first feature values and a second feature value of the second feature values.

26. The training method of claim 21, wherein the determining of the first threshold range comprises:

determining the first threshold range based on a distance between a first feature value of first feature values and a second feature value of the second feature values.

27. The training method of claim 21, wherein the training of the recognizer comprises:

training the recognizer based on an output of the first layer, an output of the second layer and a label value of a training sample, the output of the first layer being in response to an input of the training sample, the output of the second layer being in response to the input of the training sample.

28. The training method of claim 21, wherein the training of the recognizer comprises:

training the recognizer based on a first feature value output by a first network, a second feature value and a label of a training sample, the first network receives an output of the first layer, the second feature value output by a second network that receives an output of the second layer.

29. An authentication apparatus comprising:

at least one processor configured to execute computer-readable instructions to cause the authentication apparatus to,

output at least one feature value of a first object in an input image using a first layer among a plurality of layers of a deep learning (DL) network included in a recognizer, and

determine whether the first object is the same as a second object in a registered image by

determining a similarity or a distance between the at least one feature value of the input image output using the first layer and a feature value of the registered image corresponding to the first layer,

determining whether the determined similarity or distance is within a threshold range corresponding to the first layer of the layers in the recognizer, the threshold range corresponding to the first layer of the layers in the recognizer being different than a threshold range corresponding to a second layer of the layers in the recognizer, and

authenticating the input image based on whether the similarity or the distance is within the threshold range corresponding to the first layer of the layers in the recognizer.

30. The authentication apparatus of claim 29, wherein the at least one processor is further configured to execute the computer-readable instructions to cause the authentication apparatus to determine whether to accept the input image or reject the input image.

31. The authentication apparatus of claim 29, wherein the first layer belongs to a plurality of intermediate layers among the plurality of layers.

32. The authentication apparatus of claim 29, wherein the first layer belongs to a plurality of intermediate layers, and

the at least one processor is further configured to execute the computer-readable instructions to cause the authentication apparatus to determine whether the authentication of the input image succeeds based on feature values of the input image and feature values of the registered image, the feature values of the input image output by intermediate layers, respectively, and the feature values of the registered image corresponding to the intermediate layers, respectively.

33. The authentication apparatus of claim 29, wherein the at least one processor is further configured to execute the computer-readable instructions to cause the authentication apparatus to,

determine the similarity between the at least one feature value of the input image and the at least one feature value of the registered image, and

perform the authentication based on the determined similarity.

34. The authentication apparatus of claim 29, wherein the at least one processor is further configured to execute the computer-readable instructions to cause the authentication apparatus to,

determine the distance between the at least one feature value of the input image and the at least one feature value of the registered image, and

perform the authentication based on the determined distance.

35. The authentication apparatus of claim 29, wherein the at least one processor is further configured to execute the computer-readable instructions to cause the authentication apparatus to apply the registered image to the recognizer and to store at least one intermediate feature value output by at least one intermediate layer and a final layer feature value output by a final layer as feature values of the registered image.